Cognition has released SWE-1.7, their latest AI model designed for software engineering tasks. This model reportedly achieves "frontier-level intelligence" while operating at a significantly reduced cost compared to previous iterations. The improvements stem from enhancements across their reinforcement learning (RL) pipeline, including optimized infrastructure, more stable training processes, higher-quality data, and novel techniques for handling long-horizon tasks.

Built upon a Kimi K2.7 base, SWE-1.7 demonstrates substantial gains through its proprietary training, challenging the notion of a "post-training ceiling" for AI capabilities. The model is particularly suited for asynchronous tasks that span extended durations, a common requirement in complex software development. Cognition highlights SWE-1.7's performance on agentic coding benchmarks, showing a 42.3% pass rate on Frontier Code 1.1 and 81.5% on Terminal-Bench 2.1, surpassing several other leading models.

Key training innovations include methods to preserve entropy and stabilize learning, such as top-p sampling replay to prevent model collapse and maintain exploration. The team also implemented multi-cluster training across three continents to overcome compute constraints, ensuring fault tolerance and efficient weight updates. For long-horizon tasks, SWE-1.7 uses a self-compaction mechanism to summarize its working state, effectively extending its context beyond the raw window. SWE-1.7 is now available through Cognition's Devin platform.